v0.6.9.7
Two robustness fixes; the PoM walk and all consensus code are byte-identical to v0.6.9.5/.6.
Shutdown: no more crash-loop when a worker is force-killed (upstream 13d9515)
- The force-shutdown signal handler is now
extern "C-unwind": force-killing one genuinely frozen worker unwinds just that thread instead of aborting the whole miner — on HiveOS the abort turned into a restart crash-loop, worst during an OPoI inference/model reload. - The mining loop now checks for a pending shutdown or newer job before starting a multi-second resident-model reload (after an inference evicted the walk), on both the CUDA and OpenCL paths.
- (This build already had a 30s polled shutdown grace, so upstream's grace-period bump was already covered.)
Fix: ~100% "Low difficulty share" on pools/proxies that don't send daa_score
On stratum endpoints that send plain Short job notifies (no daa on the wire), the miner never crossed the H3 gate internally and kept mining unsalted pre-H3 folds — every post-H3 verifier then rejected essentially every share as "low difficulty". The internal daa floor now includes the H3 frontier once the pool declares PoM active. Suprnova was never affected (it sends real daa_scores); this hits third-party pools and stratum proxies.
Assets
NVIDIA Linux: 12 tarballs (modern sm_75+ / legacy sm_70+ incl. Tesla V100 & CMP 100-210 / pascal sm_60) + per-line SHA256SUMS. Windows zips and the macOS arm64 tarball are attached by CI. AMD Linux packages follow separately.
Package versions: CUSTOM_VERSION 0.6.9.7 / BUILD 25.